Synthetic Respondents for Busy People

Synthetic respondents are AI-generated answers that stand in for real survey participants. You describe an audience — "price-sensitive parents of toddlers", say — and a language model produces responses as if it were those people. The appeal is obvious: results in seconds, no recruitment, no incentive costs, no waiting for fieldwork.

Where this genuinely helps is rehearsal. Before you send a survey to thousands of real people, you can have AI "take" it first to surface problems: a question that reads as leading, a scale missing an option, a branch that loops back on itself, an instruction that makes no sense. This kind of dry run is cheap, fast, and catches mistakes that would otherwise waste real responses. Synthetic answers can also help you draft hypotheses worth testing, or stress-test an analysis plan before any data arrives.

The hard limit is that synthetic respondents are not evidence about the real world. A language model does not hold opinions or live with the consequences of a choice; it predicts plausible text based on its training data. Ask it which of two products people prefer and it will give you a confident answer that may be pure pattern-matching, often skewed toward whatever is common online and blind to your specific customers. It can manufacture a consensus that does not exist.

So the rule is simple. Use synthetic respondents to improve the instrument and rehearse the study — then field it for real and collect answers from actual people. The synthetic run makes your real survey better; it never replaces it.

A synthetic respondent is an AI model standing in for a human survey participant. Rather than recruiting people and collecting their answers, you prompt a large language model to simulate the responses of a described audience. The idea has spread quickly because it promises what every researcher wants and rarely gets: instant, almost-free "data". The trick is knowing which of those quotation marks to take seriously.

Where synthetic respondents earn their place

The honest, valuable uses are about preparation, not conclusions. The strongest is pre-testing. Before a survey goes to real people, you can have AI run through it and flag the things that quietly ruin studies: leading or double-barrelled questions, missing answer options, confusing logic, dead ends in the flow. Catching these before fieldwork protects your real responses from being wasted on a broken instrument. Synthetic runs are also useful for generating hypotheses to test properly later, for rehearsing an analysis plan so your spreadsheet is ready before the data lands, and for rough, clearly-labelled illustrations when no real data exists yet.

Where they quietly mislead

The fundamental problem is that a language model has no opinions, experiences, or stakes. It produces the most statistically plausible text, which is not the same as what your customers actually think. That gap creates several specific traps. It regresses to the mean of its training data, so niche, local, or recent realities get flattened into a generic internet-average view. It carries the biases of that data, often under-representing exactly the groups you most need to hear. It is sycophantic, tending to agree with the framing of your prompt, so a leading question yields a leading answer with no friction. And most dangerously, it is confident: it will hand you a clean, decisive "62% prefer option A" that looks like a finding but is closer to a hallucination dressed as data.

The line you should not cross

Synthetic respondents can rehearse a study; they cannot be the study. The moment you present simulated answers as evidence of what real people believe — to a client, to leadership, in a decision that moves money — you have crossed from a useful tool into a fabricated result. Real research derives its authority from contact with reality: actual people, with actual experiences, answering on their own behalf. Remove that, and you have a well-written guess. Used well, the workflow is a loop: rehearse with synthetic respondents to sharpen the instrument, then field the real survey and collect answers from genuine participants — and let those real answers, not the simulated ones, drive the decision.

Common Misconceptions

Common mistakes

The biggest mistake is laundering a guess into a number: running a synthetic panel, getting a tidy percentage, and presenting it as if real people said it. The second is using synthetic respondents on exactly the questions they are worst at — novel products, local markets, emotional or sensitive topics — where there is no reliable pattern to draw on and the model simply invents a confident answer. The third, less obvious, is skipping the real study entirely because the synthetic one was so easy, and never noticing that you stopped doing research and started consulting a mirror. Treated as a rehearsal tool, synthetic respondents make real research faster and sharper. Treated as a shortcut around real people, they quietly replace evidence with fiction.